Second model review works best when it reviews, not rewrites

A note on using a second model as a risk reviewer rather than a second author.

Using a second model is useful, but only if the role is narrow. If the first model writes a patch and the second model rewrites the whole thing from another style preference, the workflow becomes slower without becoming safer. The better pattern is to ask for risks: missing tests, hidden assumptions, broad changes, stale claims, or mismatch with acceptance criteria. The reviewer should return actionable findings, not a parallel implementation unless there is a real blocker. That turns model diversity into a review tool instead of model ping-pong.